모방학습 및 언어모델 기반 조립기술 개발

Abstract

Assembly is widely used in many different manufacturing applications. However, assembly is a challenging task for robot manipulators as it involves complex manipulation and interaction with environments. Many studies have been conducted for robotic assembly, and imitation learning is one of the most promising solutions as it allows robot manipulators to effectively follow demonstrated assembly trajectories. Furthermore, deep learning-based imitation learning allows highly adaptable and scalable imitation strategies. However, most of studies on imitation learning focused on household tasks, such as laundry or cleaning, and not many studies have been devoted on the application of imitation learning in manufacturing domain, such as assembly. In this study, a deep learning-based imitation learning algorithm for assembly tasks is proposed. The proposed algorithm utilizes multi-modal information to imitate complex assembly trajectories. Furthermore, to enable intuitive control of robot manipulators, large language model and language conditioned-policy are used so that users can control robot manipulators through natural language. The feasibility and performance of the proposed algorithm is verified through peg-in-hole experiments.